npm package discovery and stats viewer.

Discover Tips

  • General search

    [free text search, go nuts!]

  • Package details

    pkg:[package-name]

  • User packages

    @[username]

Sponsor

Optimize Toolset

I’ve always been into building performant and accessible sites, but lately I’ve been taking it extremely seriously. So much so that I’ve been building a tool to help me optimize and monitor the sites that I build to make sure that I’m making an attempt to offer the best experience to those who visit them. If you’re into performant, accessible and SEO friendly sites, you might like it too! You can check it out at Optimize Toolset.

About

Hi, 👋, I’m Ryan Hefner  and I built this site for me, and you! The goal of this site was to provide an easy way for me to check the stats on my npm packages, both for prioritizing issues and updates, and to give me a little kick in the pants to keep up on stuff.

As I was building it, I realized that I was actually using the tool to build the tool, and figured I might as well put this out there and hopefully others will find it to be a fast and useful way to search and browse npm packages as I have.

If you’re interested in other things I’m working on, follow me on Twitter or check out the open source projects I’ve been publishing on GitHub.

I am also working on a Twitter bot for this site to tweet the most popular, newest, random packages from npm. Please follow that account now and it will start sending out packages soon–ish.

Open Software & Tools

This site wouldn’t be possible without the immense generosity and tireless efforts from the people who make contributions to the world and share their work via open source initiatives. Thank you 🙏

© 2026 – Pkg Stats / Ryan Hefner

@aa2246740/pi-fusion

v0.2.1

Published

DRACO-verified model fusion for Pi agents

Readme

Pi Fusion

经过 DRACO 评分验证的 Pi agents 模型融合。

Pi Fusion 会把一个 prompt 分发给一组独立模型,让它们并行工作、可选地收集 evidence,再由 judge model 综合、验证并输出最终答案,整个过程发生在 Pi 里。

它的重点不只是“多跑几个模型再合并文本”。Pi Fusion 已经做过 DRACO full10 评分验证:生成阶段只看 prompt-only cases,rubric 只在生成结束后的 scorer 阶段加载,10 个 case 全部完成,并提供 sanitized public benchmark summary。

它受 hosted Fusion-style APIs 启发,但实现为 Pi extension:你选择模型,保留 artifacts,并且可以把 fusion layer 接到现有 Pi 工具和上下文中。

官网:https://aa2246740.github.io/pi-fusion/

文档:English | 中文 | Benchmarks | 中文基准说明

这个仓库为什么突出

很多 fusion demo 停在“运行多个模型并合并答案”。Pi Fusion 的定位是有评分验证的 fusion engine:

| 证据点 | 公开 claim | 为什么重要 | | --- | --- | --- | | DRACO full10 tested | 完成 10-case benchmark protocol | 这是有评分的项目,不只是 demo | | Scorer-only rubric access | rubric 只在 generation sealed 后加载 | 生成器看不到 private scoring criteria | | 对比 Fusion API | 73.80 vs reported Fusion API 69.00 | claim 绑定明确 benchmark result | | 对比 budget baseline | 73.80 vs reported Fusion API budget 64.70 | 相对 budget baseline 高 +9.10 | | 最新 validation 稳定性 | 10/10 cases,0 judge failures | 完整 protocol 跑通 | | 本地 artifact trail | answers、evidence summaries、tokens、costs 留在本地 | 用户可以检查运行过程 |

它能做什么

Pi Fusion 为一个模型答案不够用的任务提供通用 fusion layer:

  • 多个 participant models 并行独立回答
  • judge synthesis 和 contradiction analysis
  • 可选 verification/revision loop
  • 可选 web/evidence tools
  • 用于确定性计算的 sandboxed bash
  • 每个 participant 独立的 workspace sandbox 和读写工具
  • model fallback 和 retry policy
  • artifacts、evidence summaries、token usage 和 cost reporting
  • Pi 原生命令和配置

适合研究、规划、架构决策、调试假设、代码/设计评审、产品比较、写作、文档综合,以及其他需要更高可靠性的高风险问题。

安装

Pi Fusion 是 Pi package/extension。

pi install git:https://github.com/aa2246740/pi-fusion@main

npm package 名称:

npm install @aa2246740/pi-fusion

本地开发:

cd pi-fusion
npm install
pi -e ./index.ts

Pi package discovery 在 package.json 中声明:

{
  "pi": {
    "extensions": ["./index.ts"]
  }
}

命令

/pi-fusion-config       配置 participants、judge、tools 和 evidence backend
/pi-fusion-doctor       诊断模型和 evidence backend 可用性
/pi-fusion <prompt>     运行 fusion panel

示例:

/pi-fusion Should we migrate this module to a plugin architecture?
/pi-fusion --fast Summarize the tradeoffs of these three implementation plans.
/pi-fusion --quality Compare the vendors and cite current sources.

配置

运行:

/pi-fusion-config

你会选择:

  1. participant models
  2. judge model
  3. 从可选模型列表中选择 fallback models
  4. web/evidence policy
  5. evidence backend mode,通常在运行时 auto-detect
  6. sandboxed bash policy
  7. monitor 和确认默认值

配置保存在:

~/.pi/agent/pi-fusion/config.json

最小配置可以不启用 web retrieval:

{
  "participants": [
    { "model": "openai/gpt-4.1" },
    { "model": "anthropic/claude-sonnet-4-5" }
  ],
  "judge": { "model": "openai/gpt-4.1" },
  "defaultFallbacks": [],
  "webPolicy": "optional",
  "toolPolicy": { "bash": "sandboxed" },
  "monitorDefault": false,
  "confirmBeforeRun": true
}

Evidence layer

Pi Fusion 不是 scraper adapter,但它有一个可选 evidence layer,用于需要现时信息或来源支撑的任务。

如果没有配置或自动发现 evidence backend,Pi Fusion 仍然会作为 model-fusion engine 运行。如果发现兼容的 evidence backend,participant 和 judge models 可以使用 web_search 和 web_fetch tools。

运行时,Pi Fusion 会从常见 Pi/MCP 配置位置自动检测兼容的本地 MCP search/fetch servers。如果没有发现 backend,而 prompt 看起来需要 sources,Pi Fusion 会询问你是继续不带 web evidence、取消运行,还是为本次运行添加用户提供的上下文/evidence notes。

MCP evidence backend

MCP 是第一种支持的 connector type。Pi Fusion 不要求也不内置 unified-search;任何兼容 MCP search/fetch server 都可以使用。多数用户可以把 evidence backend 留给 auto-detect。高级用户可以固定一个 explicit backend:

{
  "webPolicy": "optional",
  "webBackend": {
    "type": "mcp",
    "serverName": "my-search",
    "searchServerName": "my-search",
    "searchTool": "web_search",
    "fetchServerName": "my-reader",
    "fetchTool": "web_fetch",
    "fetchFallback": "off",
    "maxResults": 5
  }
}

Expected search input:

{ "query": "..." }

Pi Fusion 可以解析常见 search result shapes,例如带有 title、url/link 和 snippet 字段的 organic[]、results[] 或 items[]。

Expected fetch input:

{ "url": "https://example.com" }

Pi Fusion 可以解析带有 content、text 或 markdown 的常见 fetch shapes。

Internal/private backends

你可以在 MCP interface 后面使用 private MCP servers、company search、本地 crawlers 或 hosted search APIs。Provider-specific backend,例如 unified-search,应该被看作可选兼容实现,而不是 Pi Fusion 核心依赖。

Workspace sandboxes

面对大型项目任务时,Pi Fusion 会把当前 Pi working directory 复制成一个 Pi Fusion 管理的 baseline,并为每个 Participant Model 创建一个独立可写 sandbox。Participant 只能通过 scoped workspace_* tools 在自己的 sandbox 中 list、search、read、write、edit 文件。

Sandbox 写入不会修改真实用户 workspace。运行结束后,Pi Fusion 会把每个 participant 的 sandbox root、变更文件列表和 ChangeSet artifacts 写到 run directory,方便 judge 和用户审查具体文件级工作,而不是只看纯文字回答。

DRACO 评分验证结果

Pi Fusion 在 sealed fixed DRACO-10 validation run 上拿到 73.80,比 reported Fusion API 69.00 高 +4.80,比 reported Fusion API budget 64.70 高 +9.10。

在这个仓库里,DRACO verified 指的是公开 benchmark claim 有完整 DRACO 10-case runs、prompt-only generation、scorer-only rubric access 和 sanitized aggregate results 支撑。它不是 DRACO 官方认证,也不是说 Pi Fusion 在所有评估里超过所有 Fusion API modes。

我们在固定 10-case DRACO validation protocol 上评估了 Pi Fusion。Final answers sealed 之后,scorer 才加载 rubric 和 scoring artifacts。

Public repository 只保留 aggregate scores、methodology notes 和 protocol checks。Raw scorer files、private rubric artifacts、case prompts 和 run internals 不公开发布。

| 验证项 | 状态 | | --- | --- | | validation cases | 10/10 | | generation input | sanitized prompt-only case files | | rubric access | scorer-only,final answers sealed 后才加载 | | latest validation judge failures | 0 | | public artifact policy | 只发布 sanitized aggregate summaries |

| System / run | DRACO score | 相对 reported Fusion API | 相对 reported budget | | --- | ---: | ---: | ---: | | Pi Fusion latest sealed DRACO-10 validation | 73.80 | +4.80 | +9.10 | | Reported Fusion API headline result | 69.00 | - | +4.30 | | Reported Fusion API budget baseline | 64.70 | -4.30 | - |

Latest validation 完成了全部 10 个 cases,judge failures 为 0。

生成阶段使用 sanitized prompt-only case files。Benchmark answer/rubric/scoring artifacts 在生成阶段不提供给 Pi Fusion,只在生成后由 scorer 使用。

更多方法论和 sanitized results 见 docs/benchmarks.zh-CN.md。

开发

npm run check
npm test

安全

Pi extensions 会以你的本地权限运行。安装第三方 package 前请先审查源码。Pi 本身不提供进程内安全 sandbox。Pi Fusion 因此把 workspace sandbox 当作产品层隔离:文件先复制到 participant 独立目录,participant 写入只留在那里,把 ChangeSet 应用回真实 workspace 是独立于 Fusion Run 的步骤。

Pi Fusion 的 bash tool 是 sandboxed,面向确定性计算,而不是 arbitrary host access。

License

MIT